llama : add Mixtral support (#4406)

* convert : support Mixtral as LLAMA arch

* convert : fix n_ff typo

* llama : model loading

* ggml : sync latest ggml_mul_mat_id

* llama : update graph to support MoE

* llama : fix cur -> cur_expert

* llama : first working version

* llama : fix expert weighting in the FFN

* ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only)

* ggml : add n_as argument to ggml_mul_mat_id

* ggml : fix ggml_get_rows to take into account ne02 / ne11

* metal : add more general support for ggml_get_rows + tests

* llama : add basic support for offloading moe with CUDA

* metal : add/mul/div use general kernel when src1 not cont

* metal : reduce the kernel launches for ggml_mul_mat_id

* ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D

* ggml : update get_rows f16 and q

* cuda : support non-contiguous src1 in get_rows

* llama : offload missing ffn_moe_silu

* metal : fix ggml_get_rows to work with non-cont src1

* metal : add indirect mat-vec kernels for all quantization types

* llama : do not quantize expert gating tensors

* llama : add n_expert and n_expert_used to hparams + change quants

* test-backend-ops : add moe test

* cuda : fix get_rows when ncols is odd

* convert : determine n_ctx correctly

* metal : fix ggml_mul_mat_id for F32

* test-backend-ops : make experts more evenly probable (test_moe)

* test-backend-ops : cleanup, add moe test for batches

* test-backend-ops : add cpy from f32 -> all types test

* test-backend-ops : fix dequantize block offset

* llama : fix hard-coded number of experts

* test-backend-ops : simplify and disable slow tests to avoid CI timeout

* test-backend-ops : disable MOE test with thread sanitizer

* cuda : fix mul_mat_id with multi gpu

* convert : use 1e6 rope_freq_base for mixtral

* convert : fix style

* convert : support safetensors format

* gguf-py : bump version

* metal : add cpy f16 -> f32 kernel

* metal : fix binary ops for ne10 % 4 != 0

* test-backend-ops : add one more sum_rows test

* ggml : do not use BLAS with ggml_mul_mat_id

* convert-hf : support for mixtral-instruct (#4428)

* convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct

* convert : use sentencepiece tokenizer for Mixtral-instruct

* convert : make flake8 happy

* metal : fix soft_max kernels

ref: ggerganov/ggml@1914017

* metal : limit kernels to not use more than the allowed threads

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: Radek Pilar <github@mrkva.eu>

Makefile

@@ -399,6 +399,11 @@ ifdef LLAMA_CUBLAS
 	MK_LDFLAGS   += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib
 	OBJS         += ggml-cuda.o
 	NVCCFLAGS = --forward-unknown-to-host-compiler -use_fast_math
+
+ifdef LLAMA_DEBUG
+	NVCCFLAGS += -lineinfo
+endif
+
 ifdef LLAMA_CUDA_NVCC
 	NVCC = $(LLAMA_CUDA_NVCC)
 else

convert-hf-to-gguf.py

@@ -77,8 +77,18 @@ def set_gguf_parameters(self):
             self.gguf_writer.add_embedding_length(n_embd)
         if (n_ff := self.hparams.get("intermediate_size")) is not None:
             self.gguf_writer.add_feed_forward_length(n_ff)
-        if (n_head := self.hparams.get("num_attention_head")) is not None:
+        if (n_head := self.hparams.get("num_attention_heads")) is not None:
             self.gguf_writer.add_head_count(n_head)
+        if (n_head_kv := self.hparams.get("num_key_value_heads")) is not None:
+            self.gguf_writer.add_head_count_kv(n_head_kv)
+
+        if (n_rms_eps := self.hparams.get("rms_norm_eps")) is not None:
+            self.gguf_writer.add_layer_norm_rms_eps(n_rms_eps)
+        if (n_experts := self.hparams.get("num_local_experts")) is not None:
+            self.gguf_writer.add_expert_count(n_experts)
+        if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
+            self.gguf_writer.add_expert_used_count(n_experts_used)
+
         self.gguf_writer.add_parallel_residual(self.hparams.get("use_parallel_residual", True))
 
     def write_tensors(self):
@@ -170,6 +180,8 @@ def from_model_architecture(model_architecture):
             return StableLMModel
         if model_architecture == "QWenLMHeadModel":
             return QwenModel
+        if model_architecture == "MixtralForCausalLM":
+            return MixtralModel
         return Model
 
     def _is_model_safetensors(self) -> bool:
@@ -207,6 +219,8 @@ def _get_model_architecture(self) -> gguf.MODEL_ARCH:
             return gguf.MODEL_ARCH.STABLELM
         if arch == "QWenLMHeadModel":
             return gguf.MODEL_ARCH.QWEN
+        if arch == "MixtralForCausalLM":
+            return gguf.MODEL_ARCH.LLAMA
 
         raise NotImplementedError(f'Architecture "{arch}" not supported!')
 
@@ -837,6 +851,11 @@ def set_gguf_parameters(self):
         self.gguf_writer.add_layer_norm_eps(1e-5)
 
 
+class MixtralModel(Model):
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+
+
 class QwenModel(Model):
     @staticmethod
     def token_bytes_to_string(b):

convert.py

@@ -42,6 +42,7 @@
 ARCH = gguf.MODEL_ARCH.LLAMA
 
 DEFAULT_CONCURRENCY = 8
+
 #
 # data types
 #
@@ -62,10 +63,10 @@ class UnquantizedDataType(DataType):
     pass
 
 
-DT_F16  = UnquantizedDataType('F16', dtype = np.dtype(np.float16), valid_conversions = ['F32', 'Q8_0'])
-DT_F32  = UnquantizedDataType('F32', dtype = np.dtype(np.float32), valid_conversions = ['F16', 'Q8_0'])
-DT_I32  = UnquantizedDataType('I32', dtype = np.dtype(np.int16), valid_conversions = [])
-DT_BF16 = UnquantizedDataType('BF16', dtype = np.dtype(np.uint16), valid_conversions = ['F32', 'F16', 'Q8_0'])
+DT_F16  = UnquantizedDataType('F16',  dtype = np.dtype(np.float16), valid_conversions = ['F32', 'Q8_0'])
+DT_F32  = UnquantizedDataType('F32',  dtype = np.dtype(np.float32), valid_conversions = ['F16', 'Q8_0'])
+DT_I32  = UnquantizedDataType('I32',  dtype = np.dtype(np.int16),   valid_conversions = [])
+DT_BF16 = UnquantizedDataType('BF16', dtype = np.dtype(np.uint16),  valid_conversions = ['F32', 'F16', 'Q8_0'])
 
 
 @dataclass(frozen=True)
@@ -151,14 +152,16 @@ def type_for_tensor(self, name: str, tensor: LazyTensor) -> DataType:
 
 @dataclass
 class Params:
-    n_vocab:    int
-    n_embd:     int
-    n_layer:    int
-    n_ctx:      int
-    n_ff:       int
-    n_head:     int
-    n_head_kv:  int
-    f_norm_eps: float
+    n_vocab:        int
+    n_embd:         int
+    n_layer:        int
+    n_ctx:          int
+    n_ff:           int
+    n_head:         int
+    n_head_kv:      int
+    n_experts:      int | None = None
+    n_experts_used: int | None = None
+    f_norm_eps:     float | None = None
 
     rope_scaling_type: gguf.RopeScalingType | None = None
     f_rope_freq_base: float | None = None
@@ -233,6 +236,13 @@ def loadHFTransformerJson(model: LazyModel, config_path: Path) -> Params:
             raise Exception("failed to guess 'n_ctx'. This model is unknown or unsupported.\n"
                             "Suggestion: provide 'config.json' of the model in the same directory containing model files.")
 
+        n_experts      = None
+        n_experts_used = None
+
+        if "num_local_experts" in config:
+            n_experts = config["num_local_experts"]
+            n_experts_used = config["num_experts_per_tok"]
+
         return Params(
             n_vocab           = config["vocab_size"],
             n_embd            = config["hidden_size"],
@@ -241,6 +251,8 @@ def loadHFTransformerJson(model: LazyModel, config_path: Path) -> Params:
             n_ff              = config["intermediate_size"],
             n_head            = (n_head := config["num_attention_heads"]),
             n_head_kv         = config.get("num_key_value_heads", n_head),
+            n_experts         = n_experts,
+            n_experts_used    = n_experts_used,
             f_norm_eps        = config["rms_norm_eps"],
             f_rope_freq_base  = config.get("rope_theta"),
             rope_scaling_type = rope_scaling_type,
@@ -255,8 +267,15 @@ def loadHFTransformerJson(model: LazyModel, config_path: Path) -> Params:
     def loadOriginalParamsJson(model: LazyModel, config_path: Path) -> Params:
         config = json.load(open(config_path))
 
+        n_experts      = None
+        n_experts_used = None
+        f_rope_freq_base = None
+
         # hack to determine LLaMA v1 vs v2 vs CodeLlama
-        if config.get("rope_theta") == 1000000:
+        if config.get("moe"):
+            # Mixtral
+            n_ctx = 32768
+        elif config.get("rope_theta") == 1000000:
             # CodeLlama
             n_ctx = 16384
         elif config["norm_eps"] == 1e-05:
@@ -266,16 +285,27 @@ def loadOriginalParamsJson(model: LazyModel, config_path: Path) -> Params:
             # LLaMA v1
             n_ctx = 2048
 
+        if "layers.0.feed_forward.w1.weight" in model:
+            n_ff = model["layers.0.feed_forward.w1.weight"].shape[0]
+
+        if config.get("moe"):
+            n_ff = model["layers.0.feed_forward.experts.0.w1.weight"].shape[0]
+            n_experts      = config["moe"]["num_experts"]
+            n_experts_used = config["moe"]["num_experts_per_tok"]
+            f_rope_freq_base = 1e6
+
         return Params(
             n_vocab          = model["tok_embeddings.weight"].shape[0],
             n_embd           = config["dim"],
             n_layer          = config["n_layers"],
             n_ctx            = n_ctx,
-            n_ff             = model["layers.0.feed_forward.w1.weight"].shape[0],
+            n_ff             = n_ff,
             n_head           = (n_head := config["n_heads"]),
             n_head_kv        = config.get("n_kv_heads", n_head),
+            n_experts        = n_experts,
+            n_experts_used   = n_experts_used,
             f_norm_eps       = config["norm_eps"],
-            f_rope_freq_base = config.get("rope_theta"),
+            f_rope_freq_base = config.get("rope_theta", f_rope_freq_base),
         )
 
     @staticmethod
@@ -832,7 +862,17 @@ def add_meta_arch(self, params: Params) -> None:
         self.gguf.add_rope_dimension_count(params.n_embd // params.n_head)
         self.gguf.add_head_count          (params.n_head)
         self.gguf.add_head_count_kv       (params.n_head_kv)
-        self.gguf.add_layer_norm_rms_eps  (params.f_norm_eps)
+
+        if params.n_experts:
+            self.gguf.add_expert_count(params.n_experts)
+
+        if params.n_experts_used:
+            self.gguf.add_expert_used_count(params.n_experts_used)
+
+        if params.f_norm_eps:
+            self.gguf.add_layer_norm_rms_eps(params.f_norm_eps)
+        else:
+            raise ValueError('f_norm_eps is None')
 
         if params.f_rope_freq_base is not None:
             self.gguf.add_rope_freq_base(params.f_rope_freq_base)
@@ -956,7 +996,7 @@ def write_all(fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyM
 
 
 def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileType:
-    wq_type = model[gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ATTN_Q].format(bid=0) +".weight"].data_type
+    wq_type = model[gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ATTN_Q].format(bid=0) + ".weight"].data_type
 
     if output_type_str == "f32" or (output_type_str is None and wq_type == DT_F32):
         return GGMLFileType.AllF32